Google Edited Its 2023 Gemini Promise on Oct 16: Connected Gmail and Drive Data Can Now Reach Human Reviewers
When Google launched Gemini’s Workspace extensions in 2023, its responsible-AI post said personal content from Gmail, Docs and Drive “is not seen by human reviewers.” That sentence now carries a note dated 16 October 2026 saying data connected through newer features can be used to train generative AI models and “can also be subject to human review in limited circumstances.”
The edit surfaced on 9 November through a GitHub repository titled “Documentation of Gemini File Leaking Behavior,” posted by an account with a single repository. It reached the Hacker News front page and passed 110 points within hours.
What Google’s own page says
The Gemini help page “About personalization with Connected Apps” lists what can be connected: Google Workspace (Gmail, Drive, Docs, Sheets, Calendar, Meet and others), Search services, Contacts, Wallet, Photos and YouTube. It states that data from those apps is used to personalise Gemini and to “improve Google services, including training generative AI models, for everyone,” and that Connected Apps data “is used and human reviewed” under the Gemini Apps Privacy Notice.
Secondary coverage quoting the same page reports further detail: reviewers include trained service providers, the reviewed subset covers summaries, excerpts and inferences from relevant emails and files, and where a message is short or especially relevant “the summary may just be the email or file.” Google reportedly says data is disconnected from the account and has personal information reduced before review, except in abuse cases, and warns users not to connect apps holding confidential information they would not want a reviewer to see. Reviewed chats are reportedly retained for up to three years.
The part that is not confirmed
The repository makes a second claim: that consumer Gemini with Connected Apps on will call Gmail or Drive tools on its own when a user mentions a personal topic, with no @Drive or @Gmail tag. The author describes four outcomes from the same prompt:
- No retrieval
- A metadata or summary reply, such as “3 emails about your flight”
- A 2 to 4 sentence verbatim quote, often with account details
- Full ingestion of the entire email, headers included
His test case was a single internet-provider bill and the prompt “I wonder if I’m paying too much for internet service.” He says all four outcomes appeared across runs, and that Google’s automated scrubbing cannot de-identify narrative data like account numbers and confirmation numbers. Those results are one anonymous author’s, are probabilistic by his own account, and have not been independently reproduced. Google had not publicly responded to the allegations at the time of reporting.
Who is affected
- Personal Google accounts aged 18 or over with Personal Intelligence on and Google Workspace connected
- Anyone who submits feedback on a chat that touched connected content: even with Keep Activity off, reporting indicates Google can take the last 24 hours of chats for review
- Work and school accounts are described in Google’s Workspace privacy hub as not human reviewed or used for training without permission, though a Drive help page on the Gemini side panel is worded more loosely
Opting out
Disconnect Google Workspace under Settings, Personal Intelligence, Connected Apps. Turning off Keep Activity also makes Workspace unavailable in Gemini. Neither step deletes existing Gemini Apps Activity or content already reviewed, and Google still holds chats for 72 hours for safety review. Temporary chats are described as exempt from improvement review.
The confirmed change is narrow but real: a 2023 absolute promise became a conditional one, in an edit made to the original post rather than a new announcement.